feat(poster): present STAMPED experiment and retained study results

.pre-commit-config.yaml

@@ -7,4 +7,4 @@ repos:
   rev: v0.9.1
   hooks:
   - id: snapper
-    files: ^2026-usrse/(stamped-abcd-poster-draft|orinoco-lite/README)\.md$
+    files: ^2026-usrse/(stamped-abcd-poster-draft\.md|orinoco-lite/README\.md|stamped-ai-study/(?!logs/).*\.md)$

2026-usrse/README.md

@@ -3,6 +3,8 @@ This is folder to collect ideas/drafts for submissions for https://us-rse.org/us
 The Orinoco Lite poster is available as [editable poster](orinoco-lite_poster.html), with [editing, PDF export and release notes](orinoco-lite/README.md).
 Its [submitted abstract](orinoco-lite-poster-draft.md) remains alongside it.
 
+The [STAMPED in Practice poster](stamped-ai-study/README.md) compares research-skill assistance across three rounds, with editable HTML, a verified PDF and links to the research evidence.
+
 ### Meeting theme etc
 
 That meeting is about/for Software Research Engineers. And this particular meeting has following

2026-usrse/stamped-ai-study/.gitattributes

@@ -0,0 +1,8 @@
+# Module-scoped equivalent for its binary types; works without MagicMime.
+# Text sources remain in Git. Large SVGs retain the repository threshold.
+* annex.largefiles=((include=*.png)or(include=*.pdf)or(include=*.jpg)or(include=*.jpeg)or(include=*.gif)or(include=*.webp)or(include=*.woff)or(include=*.woff2)or(include=*.ttf)or(include=*.otf)or(include=*.zip)or(include=*.gz)or(include=*.xz)or(include=*.zst)or(include=*.mp4)or(include=*.webm)or(include=*.bin))
+*.svg annex.largefiles=(largerthan=100K)
+
+# Poster-specific image assets are distributed through the annex remote.
+assets/*.svg annex.largefiles=anything
+*.npz annex.largefiles=anything

2026-usrse/stamped-ai-study/.gitignore

@@ -0,0 +1,10 @@
+# Keep the exported poster in this module.
+!exports/stamped-ai-poster.pdf
+.pixi/
+pixi.lock
+node_modules/
+exports/html-fullsize.png
+exports/pdf-fullsize.png
+exports/pdf-qr-*.png
+exports/qr-source/*.png
+exports/layout-review.png

2026-usrse/stamped-ai-study/BUILDING.md

@@ -0,0 +1,90 @@
+# Build and verify the poster
+
+Start with the [poster introduction](README.md) for downloads and interpretation.
+This page covers editing, exact scientific sources, environment replay and verification.
+
+## Scientific sources and display copies
+
+The [benchmark figure workflow](https://github.com/leej3/stamped-in-practice/blob/codex/stamped-poster-analysis/scripts/POSTER-FIGURES.md) contains the scientific generators, frozen inputs, root Pixi environment, derived results and provenance.
+This poster retains display copies and a [source binding](figures/benchmark-source.json), including the exact benchmark revision and SHA-256 values.
+Overview links begin with guides on `main`; direct technical links open the named research branches.
+The guides explain their role, while exact source revisions remain in the provenance.
+
+The group map shows formal round-two run 09: a one-sample group test on 54 participant maps, 9,999 sign-flip permutations and whole-brain FWE correction.
+The spatial comparison uses all 48 analyses from rounds two and three, twelve per condition, on 168,285 shared tested voxels.
+Every run retains its native corrected decisions and coverage; the source repository preserves masks, scientific code, reports and capture identities.
+
+The compact outcomes table pools the retained operational assessments from all three rounds: 72 original research runs, 70 primary scientific judgments and 68 valid primary checklists.
+Native replay counts distinguish full selected-output matches from observed executions.
+The [benchmark summary](https://github.com/leej3/stamped-in-practice/tree/codex/stamped-poster-analysis/experiments/retained-round-reporting/pooled) retains all grades, denominators, partial matches and separate round summaries.
+The [outcomes](data/benchmark-outcomes.json), [checklist index](data/benchmark-checklists.json) and [source binding](data/benchmark-outcomes-source.json) identify the exact imported reports.
+The table’s linked checklist scores average each valid primary response’s met-item count divided by all 31 checklist items, matching the app’s denominator.
+Each condition mean links to its individual filled checklists, with exact original JSON responses alongside them.
+The app rounds individual scores to integers; pooled means use the unrounded fractions and display one decimal.
+Four unavailable checklists remain unscored; their retained responses are accessible with exclusion reasons.
+The complete table includes 21 observed native reruns: 13 full selected-output matches, five partial matches and three nonzero exits.
+Round-specific results and every original judgment remain retrievable.
+
+The progression chart shows each condition across the three rounds, with separate panels for scientific adequacy and mean evaluated checklist items met.
+Each point summarizes designated primary judgments from six assigned fresh workers; science has 5–6 available judgments and the checklist has 4–6 per condition and round.
+Scientific proportions have 95% Wilson intervals, conditional on independent runs with a common within-group probability.
+The checklist panel shows each available run, connected means and ± one sample standard deviation, describing between-run spread rather than uncertainty in a mean.
+Its denominator is evaluated items (met + partial + unmet), excluding unchecked and not-applicable items; this differs from the table’s all-31-item app measure.
+The chart uses the same retained judgments as the table, with exact denominators in the linked round data.
+Neither bar type accounts for missing judgments, assessor uncertainty, shared infrastructure or the adaptive protocol; these are descriptive comparisons, not causal effects.
+Only the STAMPED condition receives the revised research-skill bundle; the other conditions retain their assigned instructions.
+The baseline label is “No supplied research skills”: common task instructions, software resources and runtime system skills remain available to all conditions.
+
+The seven concise definitions follow the [STAMPED paper](https://github.com/stamped-principles/stamped-paper/blob/main/main.tex), including component and change provenance, versioned environment specifications, persistent retrieval and explicit licenses.
+A spectrum note distinguishes incremental improvement from a single compliance threshold.
+
+The poster focuses on the scientific results and skill-development experiment.
+Original reporting failures, additive extraction corrections and infrastructure diagnostics remain in the benchmark's linked reporting record.
+Future directions draw on the [study lessons](https://github.com/leej3/stamped-in-practice/blob/main/LESSONS.md): stronger calibrated judges, clear completion rules, common tool guidance, measured coordinator effort, isolation, execution capture and new scientific tasks.
+
+## Retrieve and build
+
+Run from this directory:
+
+```sh
+pixi install
+pixi run datalad get ../orinoco-lite/assets/con-logo-outlined.svg \
+  ../orinoco-lite/assets/con-ccn-dartmouth-letterhead.svg \
+  ../orinoco-lite/assets/repronim_logo.svg \
+  ../orinoco-lite/assets/orinoco-qr-con.svg assets figures exports
+pixi run setup-browser
+pixi run build
+pixi run verify
+pixi run links
+```
+
+`pixi run build` verifies the scientific display copies and outcome table against their bindings, then exports the local HTML/CSS poster. It does not regenerate scientific analyses or request live study data. Before replacing annexed exports, run `pixi run git annex unlock exports/stamped-ai-poster.pdf exports/pdf-preview.png`. The exact validated `pixi.lock` is preserved separately on [`codex/env-poster-ci-20261016`](https://github.com/con/talks/tree/codex/env-poster-ci-20261016). For an exact environment replay, retrieve its `2026-usrse/stamped-ai-study/pixi.lock` here before `pixi install --locked`.
+
+To import a new committed benchmark export:
+
+```sh
+pixi run git annex unlock figures
+pixi run python scripts/import-benchmark-figures.py --benchmark /path/to/stamped-in-practice
+pixi run python scripts/import-benchmark-outcomes.py --benchmark /path/to/stamped-in-practice
+pixi run build
+```
+
+The importer checks every source recorded in the benchmark export manifest and requires those files to match its committed revision.
+Scientific regeneration and input refresh happen through the benchmark's documented root tasks.
+The historical poster-side scientific pipeline remains in Git history; it is no longer a second source of truth.
+
+Institutional marks are reused from the ORINOCO Lite poster.
+To change QR destinations, edit `make-logo-qrs.py` and run `pixi run python make-logo-qrs.py --check-dir exports/qr-source`.
+QR regeneration additionally requires Inkscape and `rsvg-convert`; ordinary builds use the supplied QR assets.
+
+## Verification and provenance
+
+[HTML audit](exports/html-audit.json), [PDF audit](exports/pdf-audit.json), [link audit](exports/link-audit.json), [retrieval audit](exports/retrieval-audit.json) and [visual review](exports/agent-visual-review.json) describe the current export.
+The retrieval audit records fresh full-payload checks for the current PDF, preview and changed figure assets, alongside earlier checks for unchanged assets.
+Checks require one 44-inch square page, at least 24-point print text, no layout overflow, clickable PDF links, and four QR codes decoded from PDF pixels.
+These are agent checks; author review remains part of poster preparation.
+
+Earlier designs remain in the [poster history](https://github.com/con/talks/commits/codex/stamped-example-poster/2026-usrse/stamped-ai-study).
+[source-manifest.json](source-manifest.json) records the inherited layout and artwork.
+Composition and prose are CC BY-SA 4.0; scripts are Apache-2.0.
+Institutional marks and external scientific inputs retain their own rights, documented in [REUSE.toml](../../REUSE.toml) and the benchmark provenance.

2026-usrse/stamped-ai-study/README.md

@@ -0,0 +1,44 @@
+# STAMPED in Practice poster
+
+Can research guidance help AI agents do sound science and leave work others can inspect and reproduce?
+This **44 × 44 inch** poster presents a comparison of four kinds of assistance across three rounds of a shared neuroscience task.
+It shows both scientific results and how the research guidance was tested and revised.
+
+## Read the poster
+
+- [Download the current PDF](https://hub.centerforopenneuroscience.org/git-annex-p2phttp/git-annex/40d344f5-0628-494e-82e4-9fab0949ed12/key/MD5E-s908837--31099438fe11c5bd19df7b43513b8e42.pdf).
+- [View the image preview](https://hub.centerforopenneuroscience.org/git-annex-p2phttp/git-annex/40d344f5-0628-494e-82e4-9fab0949ed12/key/MD5E-s1116391--0e3f41dd0df6a3610fb5b7634c75608e.png).
+- [Read the poster and figure guide](https://github.com/leej3/stamped-in-practice/blob/main/docs/poster.md) for context before following individual research records.
+
+The PDF and preview are publicly downloadable; their contents have been checked against the reviewed local export.
+GitHub may show annexed files as pointers, so use the download links above rather than expecting those pointers to display the poster.
+
+## Follow the story
+
+The left column introduces STAMPED and the experiment.
+The centre shows one agent's brain analysis and where repeated analyses agree.
+The right compares scientific adequacy and research practices, including their progression across rounds.
+The scientific and checklist measures describe different outcomes; an improvement in one does not establish an improvement in the other.
+The plot shows 95% Wilson intervals for scientific adequacy and individual checklist scores with means ± one standard deviation.
+Each linked table score opens the condition’s individual filled checklists and original responses.
+The table uses mean app scores out of all 31 items; the plot uses the fraction of evaluated items met, a separate measure.
+
+Choose a starting point for the part you want to explore:
+
+- [Results and interpretation](https://github.com/leej3/stamped-in-practice/blob/main/docs/results.md) explains the measures, missing assessments and limits of the comparison, then leads to the underlying tables.
+- [Scientific figures](https://github.com/leej3/stamped-in-practice/blob/main/docs/poster.md#interpret-the-scientific-figures) explains what each image represents and where its inputs and plotting code live.
+- [Research skills and assessment](https://github.com/leej3/stamped-in-practice/blob/main/docs/skills.md) introduces the workflows, the independent assessor and the revisions tested in each round.
+- [Repository map](https://github.com/leej3/stamped-in-practice/blob/main/docs/repository-map.md) explains why evidence links open different branches and how to read an individual research delivery.
+
+The poster's general links and QR codes lead to these introductions.
+Links labelled **Analysis code**, **Numerical results**, **Figure code**, **Map inputs** or **Round data + denominators** open the named material directly.
+
+## Edit or reproduce the poster
+
+This directory is in the presentation repository, `con/talks`, on the `codex/stamped-example-poster` branch.
+The scientific records and figure-generation workflow belong to `leej3/stamped-in-practice`; this directory holds display copies, their source bindings and the editable [HTML](index.html) and [CSS](poster.css).
+The [build guide](BUILDING.md) covers retrieval, source imports, the exact environment and export checks.
+
+The exported poster passes checks for its page size, minimum 24-point text, links, QR codes and public downloads, plus a visual review.
+These agent checks support author review.
+The [build guide's verification records](BUILDING.md#verification-and-provenance) contain the details; the [completion record](completion-plan.md) and [poster history](https://github.com/con/talks/commits/codex/stamped-example-poster/2026-usrse/stamped-ai-study) preserve how the poster developed.

2026-usrse/stamped-ai-study/assets/stamped-qr-poster.svg

@@ -0,0 +1 @@
+../../../.git/annex/objects/7P/pg/MD5E-s71326--5efd8ccdbf4534a4377167b6e72e77a8.svg/MD5E-s71326--5efd8ccdbf4534a4377167b6e72e77a8.svg

2026-usrse/stamped-ai-study/assets/stamped-qr-skills.svg

@@ -0,0 +1 @@
+../../../.git/annex/objects/x2/Fp/MD5E-s71260--1996f2b5c3ec7072029d9ccb4c6291a5.svg/MD5E-s71260--1996f2b5c3ec7072029d9ccb4c6291a5.svg

2026-usrse/stamped-ai-study/assets/stamped-qr-study.svg

@@ -0,0 +1 @@
+../../../.git/annex/objects/jp/qZ/MD5E-s67623--fad8037a2c4eac1c81b0f96a75bb69e2.svg/MD5E-s67623--fad8037a2c4eac1c81b0f96a75bb69e2.svg